Repetition is the mother of wisdom, not wisdom itself
Four filters that ranked our memory by volume put plans and run logs on top. Measuring novelty instead put them at the bottom, where they belong. Then the criterion hit a wall: our register of known concepts was a list of filenames.
1 Nov 2026 · 2 min readWhich records in a memory are knowledge, and which are a log of knowledge being used? Four filters failed to tell them apart. Then Michal put it in one sentence:
When a record contains many entities that already exist and not one new one, it is a log, a record of how we used knowledge on the way, running the same neural paths we already have. Repetition is the mother of wisdom, but not wisdom itself.
Behind it is Prof. Antonín Rosický’s systems question: what is a source of knowledge, and what is merely a record of activity in which knowledge was used?
We had been measuring volume. The right quantity is increment.
The measurement
Walking the memory chronologically, for each file: what share of its terms had never appeared before?
| family | share of terms that are new |
|---|---|
| concepts | 69% |
| design | 65% |
| lessons | 63% |
| findings | 62% |
| plans | 56% |
| run logs | 48% |
The confound, removed
Novelty falls with position, because early files see an emptier corpus:
1st quarter 70% · 2nd 58% · 3rd 47% · 4th 64%
Compared within each quarter, the order holds. Run logs come last in the first three quarters. The signal is real and not an artefact of ordering. The first of five filters that points the right way: volume put plans and runs on top; novelty puts them at the bottom.
Why it cannot be a threshold yet
48% against 69% will not separate cleanly. Pushing further, asking which new terms later took hold (appeared in three or more files), the order flips:
runs 15% · plans 12% · concepts 6% · design 6% · findings 4% · lessons 0%
What runs introduce: component filenames, a backup command, template headings like “Written by:”. They recur because the template recurs. Lessons score zero because they are short: one sentence of hard-won truth can carry no new noun at all.
The counter measures strings, not concepts. That is the honest limit of pattern matching over prose.
The decisive test
Apply the criterion literally: compare each file against the register of known concepts, the knowledge graph.
known concepts in the register 1,025
median known concepts found per file 0 (every family, all 986 files)
Zero. Because 45% of the “concepts” were filenames and event titles, strings that never occur in the prose of other files.
The conclusion that closes the loop
The criterion is correct and currently unmeasurable. It needs exactly what distillation is meant to produce: a register of real concepts instead of a list of files. Until then, any filter for “knowledge versus log” measures the wrong thing, however clever it looks.
The measurements in this article come from these records in the running system.
~/memory/MEASURED-novelty-not-volume-the-rosicky-criterion.md— novelty, not volume: the Rosický criterion
Every counter said 1,047 memories were indexed. 653 of them were their opening 4,000 characters. 72% of the text never reached the vector store.
12 MB of vectors, a 39 GB index, zero dead rows. A nightly sync updated every row whether or not anything had changed, and a vector index never gives back space from updates. Ten million no-op writes, found by one ratio.
Our unified search tool says 'prefer this over individual tools'. Our entire research corpus was not one of its layers. On six topics that live only there, the direct search won 6 of 6. And 75% of our memory had never been retrieved even once.